Deepika Bablani

LLM post training @Apple.

Previously @IBMResearch, @CMU.

[Deepika Bablani @deepikabablani]

all models are wrong, some are useful.

remember that every time a new model drops.

[Deepika Bablani @deepikabablani]

@DimitrisPapail Well it’s bearing a lot of the load tbf… :)

[Deepika Bablani @deepikabablani]

i like green things - grass, matcha, nvidia gpus, pistachio gelato…. not necessarily in that order.

[Deepika Bablani @deepikabablani]

@tszzl If you remove the word "worship" in the first paragraph, which is a conjecture at best, there is nothing surprising there. Many organizations love the product they create, and use AI in some form for screening, performance reviews, and development today.

[Deepika Bablani @deepikabablani]

Meahwhile, someone just got asked how to find container with most water on a technical coding interview in 2026.

[Deepika Bablani @deepikabablani]

P1 creates a new repo overnight with 30k LOC. 2 days later, P2 creates a v2 complete with migration guide, 29k - / 31k + LOC in single PR. Suggests too big to review but works on few test cases, should just switch. P1 refuses to review, agrees will run few tests and merge. RIP SWE.

[Deepika Bablani @deepikabablani]

@natfriedman @jessfraz 5. Population overhang: Such a small percentage of human population is actively thinking about / working on improving AI.

[Deepika Bablani @deepikabablani]

All you need is <strikethrough>love</strikethrough> good evals.

[Deepika Bablani @deepikabablani]

TIL that mandarin has a higher information density per character than English. Interestingly, this might be one reason why LLMs switch to Mandarin for thinking occasionally.

[Deepika Bablani @deepikabablani]

To err is human, to compress divine.

[Deepika Bablani @deepikabablani]

Bc SF summer is September-October.

[Deepika Bablani @deepikabablani]

I got my first Pro subscription
Bought it at the Five and Dime
Prompted 'til my fingers bled
Was the summer of '996 #aisummer #summerof996 #backtowork

[- Elijah Muraoka - @elijahmuraoka_]

My Uber driver just reviewed and merged a PR while driving

I swear this shit only happens in SF

[Deepika Bablani @deepikabablani]

How do you know your human labelers are also not just using AI for their job? 🤔 #Synthetic #AIGeneration #Evals

[Deepika Bablani @deepikabablani]

Ok, it can if you explicitly ask it to not, but the policy is optimized for follow ups. Yet another engagement play.

[Deepika Bablani @deepikabablani]

I haven't seen a single case yet of GPT-5 not responding with a follow up question at the end. Reward hacking much? Or is it in the system prompt? #GPT5 #rlhf